Causal Machine Learning in Commodity Markets: A Framework for Oil Price Forecasting

This paper introduces a modeling framework that integrates constraint-based causal discovery with predictive algorithms for oil market analysis. The methodology first applies the PC algorithm to identify a causal graph from heterogeneous market data. This graph then informs feature selection for a LightGBM model, constraining it to causally-relevant variables. Empirical results demonstrate that this approach maintains forecasting accuracy while providing interpretability through SHAP analysis and counterfactual reasoning. The derived causal structure corroborates established economic principles, highlighting inventory dynamics and regional arbitrage as primary price drivers.